acceptodds
Under review as a conference paper at ICLR 2027

DeltaSplat: Iterative Gaussian Refinement for Pose-Free Feed-Forward 3D Gaussian Splatting

Abstract

Pose-free feed-forward 3D Gaussian Splatting (3DGS) reconstructs a scene from sparse, unposed images in a single network pass, removing the need for camera calibration and per-scene optimization. However, camera estimation errors propagate into the predicted Gaussians and compound the geometric and photometric inaccuracies of single-pass prediction. To correct these errors, we introduce DeltaSplat, a lightweight Gaussian refinement module for pose-free feed-forward 3DGS. It iteratively renders the current Gaussians at the input context views and predicts per-Gaussian updates from the resulting residuals. A 2D residual alone, however, underdetermines the 3D correction. DeltaSplat therefore conditions each update on per-pixel Plücker rays and rendered depth as a soft geometric prior. A dual-branch convolutional mixer efficiently encodes these inputs, and per-attribute heads decode the fused features into position, opacity, and color updates. The module adds only  2.2% parameters to the backbone and remains fully feed-forward at inference. On DL3DV, DeltaSplat reaches 26.64 dB PSNR in the pose-free setting, improving its state-of-the-art backbone by 1.75 dB and surpassing even baselines supplied with ground-truth cameras; consistent gains hold across 6-24 views and all camera regimes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, feed-forward reconstruction, pose-free, novel view synthesis, iterative refinement, sparse-view reconstruction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xDDdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DeltaSplat iteratively refines a feedforward Gaussian distribution using rendering residuals based on Plücker rays and rendering depth. A shared refinement module is responsible for predicting updated values for position, opacity, and color. Experimental results show that the performance of novel view synthesis are improved across various camera setups, numbers of views, and datasets.

Critical strengths
  1. The improvement in rendering performance is significant: on the DL3DV dataset, when using predicted poses and known intrinsic parameters, DeltaSplat outperforms YoNoSplat by 1.75 dB, and this improvement is consistently maintained across different numbers of views and on other datasets.

  2. Experiments on removing geometric conditions also valuable: removing ray and depth information resulted in a 1.37 dB decrease in PSNR, indicating that these components are critical to the proposed architecture.

Critical weaknesses
  1. The robustness of this method against camera uncertainties has not yet been directly verified. Whether using a simulated or a real camera, the resulting gains cannot be distinguished between error correction and a general Gaussian smoothing process.

  2. Illumination and depth data are removed jointly, it is not possible to clearly identify their respective contributions or how they complement each other.

Questions for authors
  1. Can controlled camera perturbations demonstrate whether geometric conditioning improves robustness to pose errors?

  2. Can you provide ray-only and depth-only ablations to clarify their individual contributions and complementarity?

3 comments
AuthorsOpenReview ↗

Response to Reviewer xDDd

We thank the reviewer for highlighting camera uncertainty and the individual roles of the geometric inputs. We address both concerns with additional experiments.

Q1. Does camera perturbation testing show robustness beyond generic smoothing?

We added controlled camera perturbations on 140 DL3DV scenes at 224×224, using rotation/translation errors of 1°/2%, 2°/5%, and 5°/10% with 2, 4, 6, 12, and 24 input views. DeltaSplat outperforms ReSplat in all 15 settings by 0.13-1.91 dB (Table R8). Its rendering advantage therefore persists across the tested uncertainty levels, although absolute quality decreases as perturbations grow. At the strongest perturbation (5°/10%), the gains are +1.91 dB with two views and +0.13 dB with 24 views. Thus, gains remain positive but vary with the input configuration.

To examine the role of explicit geometry, we separately ablate the conditioning inputs. Removing rays and depth reduces PSNR by 1.37 dB (Table R7), while Figure 5 shows larger, more dispersed position updates. Conditioning thus affects spatial corrections as well as output image quality. Because cameras remain fixed, the perturbation study evaluates whether Gaussian updates retain a rendering advantage when camera geometry is imperfect. Together, these results support better rendering under camera errors and geometry-dependent refinement. The full-model comparison does not isolate conditioning's contribution to robustness or establish metric 3D accuracy.

Q2. What are the individual and complementary contributions of rays and depth?

We added ray-only and depth-only variants (Table R7). Without geometry, PSNR is 25.27 dB; rays alone reach 25.93 dB and depth alone reaches 26.25 dB. Each input therefore contributes, with depth providing the larger individual gain in this setting. Both single-input variants also improve SSIM and LPIPS.

The combined model achieves 26.64 dB PSNR, 0.862 SSIM, and 0.115 LPIPS, the best result across all three metrics. Adding rays to depth improves PSNR by a further 0.39 dB; adding depth to rays provides 0.71 dB. Joint conditioning therefore improves on either signal alone.

The signals provide distinct spatial information. Plücker rays describe viewing lines derived from fixed context cameras. Rendered depth estimates where the current reconstruction lies along those lines and changes as the Gaussians are refined. A 2D residual alone does not uniquely determine a 3D correction: an image error can be reduced through position, opacity, or color changes. The two signals provide complementary spatial context for choosing these updates without camera reoptimization. These estimated geometric inputs serve as soft conditioning: they provide context for predicting corrections, without constraining updated positions to match the current rendered depth.

Tables R1–R8 are provided in the accompanying Official Comments titled "Supporting Experiments (1/2)" and "Supporting Experiments (2/2)".

Thank you for the clarifications and additional experiments. The reported ray-only and depth-only ablations address my concern about their individual contributions and support the benefit of joint conditioning.

The perturbation results provide useful evidence of improved rendering under camera errors, although they do not isolate whether geometric conditioning itself improves robustness. Please update the manuscript PDF to include the new experiments.

Overall, the response substantially addresses my concerns, and I maintain my weak accept recommendation.

AuthorsOpenReview ↗

Response to Reviewer xDDd

Thank you for your thoughtful feedback and for acknowledging that the additional ablations address your concerns. We will incorporate the new ablation and camera perturbation results into the revised manuscript.

Reviewer_X71aOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DeltaSplat, a lightweight Gaussian refinement module for pose-free feed-forward 3DGS. It iteratively renders the current Gaussians at the input context views and predicts per-Gaussian updates from the resulting residuals. A 2D residual alone underdetermines the 3D correction. DeltaSplat therefore conditions each update on per-pixel Plücker rays and rendered depth as a soft geometric prior, and per-attribute heads decode the fused features into position, opacity, and color updates. The module adds only ~2.2% parameters to the backbone and remains fully feed-forward at inference.

Critical strengths
  1. Well-motivated problem and clean, effective solution. The paper addresses a real pain point in pose-free feed-forward 3DGS: camera estimation errors propagate into the predicted Gaussians, causing systematic misalignment and ghosting. DeltaSplat tackles this directly by iteratively refining the Gaussian representation based on rendering residuals, and the geometric conditioning (Plücker ray + depth) is a sensible way to resolve the 2D residual's inherent 3D ambiguity. The result is a lightweight, plug-and-play module (~2.2% additional parameters) that delivers consistent and measurable improvements.

  2. Comprehensive and well-designed experiments. The evaluation goes beyond the main DL3DV results: cross-dataset generalization to ScanNet++ and RE10K, application to alternative backbones (NoPoSplat, AnySplat), and a thorough ablation study isolating geometric conditioning, dual-branch mixer design, and the number of refinement iterations. The per-scene analysis also shows improvements are consistent across scenes rather than driven by outliers.

Critical weaknesses
  1. The core paradigm, iterative residual-based Gaussian refinement is not new. The idea of rendering the current reconstruction, computing photometric residuals against the input views, and predicting Gaussian updates from those residuals has already been established by ReSplat and SplatFormer in the calibrated setting. DeltaSplat's main contribution is applying this existing paradigm to the pose-free setting, where camera errors introduce additional spatial ambiguity. While this is a meaningful domain shift with practical value, it does not offer a fundamentally new insight into how residual-based refinement works or should work. The geometric conditioning via Plücker rays and depth is also a natural extension to handle the 2D-to-3D ambiguity introduced by pose errors, rather than a surprising or counterintuitive design choice. The paper would benefit from a sharper articulation of what is conceptually new beyond "we took an existing refinement paradigm and applied it to a harder setting."
  2. Missing comparison with the most closely related concurrent work (GFSplat). The paper mentions GFSplat in the related work as a concurrent approach addressing pose-free spatial ambiguity using generative diffusion priors, but does not include it in the experimental comparison. Given that GFSplat targets essentially the same problem and is published in the same timeframe, a direct comparison is expected and its absence is a significant gap.
  3. No analysis of why performance saturates and degrades beyond K=3. Table 5 shows that PSNR peaks at K=3 (26.64 dB) and gradually drops for K=4–6 (26.48, 26.32, 26.17 dB). The paper describes this phenomenon empirically but does not investigate why it happens. Is the model overfitting to the input context views with more refinement iterations? Does the residual signal become ambiguous or noisy as the reconstruction converges? Without understanding the failure mode, the K=3 choice feels empirical rather than principled, and readers cannot tell whether a better design (e.g., adaptive iteration stopping, or a different refinement loss) could extend the gains further.
Questions for authors
  1. How does this compare to GFSplat? It seems like the closest concurrent work. What's the difference in approach, and how do the numbers stack up on the same benchmarks?

  2. Table 5 shows performance peaking at K=3 and dropping after that. Why does this happen? Is the model overfitting to the input views with more iterations?

  3. How much of the gain comes from the extra compute (three more forward passes) versus the refinement idea itself?

  4. In the fully uncalibrated setting, PSNR drops by about 1 dB. Is this mostly intrinsics error or extrinsics error? Does refinement still help when cameras are way off?

  5. What happens with very sparse input—like 3 or 4 views? The paper tests 6–24 views, but pose-free matters most when views are extremely few.

  6. Since cameras are fixed during refinement, how sensitive is this to initial pose accuracy? If predicted cameras are way off, will the method still recover correct 3D geometry, or will it distort Gaussians to compensate—rendering fine on input views but drifting further from true geometry?

5 comments
AuthorsOpenReview ↗

Response to Reviewer X71a

We thank the reviewer for the constructive questions on the contribution and refinement behavior. We first address the novelty concern (W1), then follow the reviewer's question order (Q1–Q6), noting the related weaknesses for each question.

W1. What is conceptually new?

DeltaSplat supplies Plücker rays and current rendered depth as explicit backbone geometry priors, helping a lightweight refiner interpret ambiguous 2D residuals as 3D corrections and release the bottleneck of the model capacity without additional prior model. This geometry-conditioned iterative design improves FF3DGS backbones across camera regimes (Tables R1, R7).

Q1 (W2). How does DeltaSplat compare with GFSplat?

We understand GFSplat to refer to GIFSplat, the concurrent diffusion-prior method discussed in our related work. GIFSplat adds a DiFiX diffusion prior; DeltaSplat reuses backbone geometry. With matched 10,000-step training, six-view YoNoSplat gains +0.88 dB versus GIFSplat/IFSplat's +0.39/+0.30 dB using predicted poses/known intrinsics (Table R1). Runtime is reported in Table R3.

Q2 (W3). Why does performance decline beyond K=3? Is it input-view overfitting?

PSNR falls from 26.64 dB at K=3 to 26.17 dB at K=6 (Table R5). Training supervises the three-step endpoint through context and target views. Longer rollouts exceed this trained horizon without a convergence guarantee; the decline does not establish input-view overfitting. Per-iteration gains shrink from +1.22 to +0.43 to +0.19 dB (Table R5), consistent with a diminishing residual signal as the reconstruction converges. Adaptive stopping or training with longer or randomized rollouts may extend the useful horizon; we will discuss this in the revision.

Q3. How much of the gain comes from extra compute?

At fixed K=3, conditioning adds 1.37 dB (Table R7), while further iterations worsen quality (Table R5), supporting design-dependent gains. Additional passes execute only the lightweight refiner, without repeating the backbone (106 ms over the 54 ms backbone at K=3; Table R5).

Q4. Intrinsic or extrinsic error? Does refinement help when cameras are far off?

At six views, predicting poses and then intrinsics introduces sequential PSNR penalties of 0.06 and 0.95 dB (Table R6). The larger loss accompanies predicted intrinsics, although pose and geometry estimates can also change. Refinement still helps in the fully uncalibrated regime, exceeding YoNoSplat by 1.16/1.41/1.49 dB at 6/12/24 views (Table R6), and remains above ReSplat under 5°/10% perturbations (Table R8).

Q5. What about very sparse inputs?

With predicted poses and intrinsics, two-/four-view gains are +0.07/+0.39 dB for YoNoSplat and +0.47/+0.79 dB for AnySplat (Table R1). Under camera perturbations, the two-view advantage over ReSplat is largest (+1.35 to +1.91 dB; Table R8). Improvements persist with few views, although the two-view YoNoSplat gain is modest and refiner rankings vary.

Q6. Can fixed cameras distort geometry?

DeltaSplat exceeds ReSplat in all 15 perturbation conditions (Table R8), and no-alignment evaluation also shows novel-view improvements (Table R4). Target-view supervision encourages generalization, but Gaussian updates can compensate for fixed-camera errors; these rendering results do not certify metric 3D accuracy.

Tables R1–R8 are provided in the accompanying Official Comments titled "Supporting Experiments (1/2)" and "Supporting Experiments (2/2)".

AuthorsOpenReview ↗

Follow-up to Reviewer X71a

Thank you for your careful review and valuable feedback. We have responded to your comments with further discussion and additional experiments, and we hope these help clarify the contributions and evaluation of our work.

We would be grateful for any feedback on our response, particularly if there are concerns that remain unresolved. Please let us know if additional explanation would be helpful; we would be happy to clarify any points during the discussion.

Thank you again for your time.

AuthorsOpenReview ↗

Follow-up to Reviewer X71a

Dear Reviewer X71a,

We sincerely thank you for your careful and detailed review, and we fully understand that this is a very busy period. As the discussion period will close in three days, we would like to kindly and respectfully follow up on our response. In our rebuttal, we humbly tried to address the points you raised:

  • Regarding the comparison with GIFSplat, the concurrent diffusion-prior method, DeltaSplat gains +0.88 dB vs. +0.39 dB for GIFSplat on six-view YoNoSplat under matched training, without relying on a diffusion prior (Tables R1, R3).
  • Regarding the source of the gains, at a fixed K=3, removing the geometric conditioning costs 1.37 dB, while additional iterations beyond K=3 do not help (Tables R5, R7), which we hope suggests that the gains come from the design rather than extra computation.
  • Regarding sparse views and camera errors, DeltaSplat outperforms ReSplat in all 15 perturbation settings, with the largest margin at two views (Table R8).

We hope these results are helpful. If any concerns remain, we would be truly grateful for even a brief comment, and we would very much welcome the opportunity to respond further or provide any additional clarification before the discussion closes.

Thank you once again for your valuable time and thoughtful feedback.

Respectfully,
The Authors

Reviewer_1E43OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paer proposes DeltaSplat to refine Gaussians for pose-free feed-forward 3DGS. It conditions updates on per-pixel Plucker rays and rendered depth and uses dual-branch convolutional mixer to update attributes.

Critical strengths
  1. This paper is well-written and easy-to-follow.
  2. The proposed method achieves better performance than evalauted baselines.
Critical weaknesses
  1. The proposed method appears to be an incremental adaptation of ReSplat-style iterative refinement to a pose-free backbone. Residual-driven Gaussian refinement has already been explored in ReSplat and GIFSplat; importantly, the latter’s IFSplat variant already performs pose-free iterative refinement without a generative prior. Camera-ray conditioning and Plücker representations also have precedents in YoNoSplat and PF3plat, respectively. arXiv Similarly, multi-scale feature processing and coarse-to-fine Gaussian reconstruction have been explored in ReSplat and HiSplat.
  2. Substantial inference-time overhead.
Questions for authors
  1. Please see the weakness section.
  2. How do DeltaSplat perform without evaluation-time pose alignment?
  3. Could DeltaSplat and ReSplat be compared using the same backbone under comparable training settings and the same evaluation protocol, reporting both rendering quality and inference time?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 1E43

We thank the reviewer for the constructive assessment. The additional experiments clarify our contribution, cross-backbone applicability, and computational cost. Since Q1 refers to the weakness section, we address W1 under Q1 and W2 (inference overhead) under Q3, which also asks for inference time.

Q1 (W1). How does DeltaSplat differ from prior refinement and ray conditioning?

DeltaSplat is a general FF3DGS refinement module addressing the geometric-inference burden of a lightweight refiner: 2D residuals alone do not specify the necessary 3D corrections, especially with estimated cameras. We expose the backbone's geometric prior through fixed-camera Plücker rays and depth rendered from the evolving Gaussians. This supplies spatial context that the small refiner would otherwise need to infer from residuals alone. A dual-branch multi-scale mixer captures broad context at reduced resolution and fine residual details at full resolution to guide position, opacity, and color updates. ReSplat and HiSplat also use multi-scale processing; here, the branches encode geometry-conditioned residuals for updates to an existing Gaussian representation.

YoNoSplat's Intrinsic Condition Embedding adds intrinsic-derived ray embeddings to image features before the decoder for initial reconstruction. PF3plat combines pose-derived Plücker features with depth-network features to predict camera-pose corrections. DeltaSplat conditions Gaussian updates while keeping cameras fixed. Compared with GIFSplat's external DiFiX diffusion prior, it reuses guidance from the backbone and current reconstruction. IFSplat, GIFSplat's prior-free variant, also performs pose-free iterative refinement; under matched training, it gains +0.30 dB on six-view YoNoSplat versus +0.88 dB for DeltaSplat (predicted poses/known intrinsics; Table R1). Removing geometric conditioning costs 1.37 dB (Table R7), while cross-backbone results demonstrate gains across pose-aware and pose-free settings (Table R1).

Q2. How does DeltaSplat perform without evaluation-time pose alignment?

PSNR and LPIPS improve in all six tested settings without alignment (Table R4; Appendix D). This alignment adjusts target cameras against held-out images and is separate from Gaussian reconstruction. Its removal shows that refinement benefits persist without that evaluation procedure.

Q3 (W2). How do DeltaSplat and ReSplat compare on the same backbone in quality and inference time?

Quality: We trained all compared refiners for 10,000 DL3DV steps and evaluated the same 140 scenes at 224×224 (Table R1). On six-view YoNoSplat with predicted poses/known intrinsics, DeltaSplat gains +0.88 dB versus ReSplat's -0.01 dB. ReSplat performs better on its own bare backbone, while DeltaSplat further improves its refined outputs. At 256×448, adding DeltaSplat after ReSplat's four iterations yields another +0.30/+0.36/+0.38 dB at 8/16/32 views (Table R2).

Inference time: Six-view refinement takes 0.16 seconds versus 0.20 for ReSplat and 1.10 for GIFSplat. At 24 views, the corresponding times are 0.84/0.49/4.30 seconds; IFSplat takes 0.70 seconds (Table R3). The quality gains thus involve higher cost than ReSplat/IFSplat at larger view counts.

Tables R1–R8 are provided in the accompanying Official Comments titled "Supporting Experiments (1/2)" and "Supporting Experiments (2/2)".

AuthorsOpenReview ↗

Follow-up to Reviewer 1E43

Thank you again for your thoughtful review and constructive comments. In our response, we have provided additional explanations and experimental results to address the questions raised in your review.

We would appreciate it if you could let us know whether these responses have addressed your concerns. If any aspect of our approach or results would benefit from further clarification, we would be happy to provide more details.

Thank you for your time and consideration.

AuthorsOpenReview ↗

Follow-up to Reviewer 1E43

Dear Reviewer 1E43,

We sincerely thank you for the time and effort you have devoted to reviewing our paper, and we fully understand that this is a very busy period. As the discussion period will close in three days, we would like to kindly and respectfully follow up on our response. In our rebuttal, we humbly tried to address the points you raised:

  • Regarding the difference from ReSplat and IFSplat, we provided a matched-training comparison on the same pose-free backbone, where DeltaSplat gains +0.88 dB vs. +0.30 dB (IFSplat) and −0.01 dB (ReSplat) at six views (Table R1), as well as an ablation showing that removing the geometric conditioning costs 1.37 dB (Table R7).
  • Regarding inference overhead, we reported refinement times (e.g., 0.16 s vs. 0.20 s for ReSplat at six views; Table R3).
  • Regarding evaluation without pose alignment, we found that DeltaSplat still improves PSNR and LPIPS in all six settings (Table R4).

We hope these results are helpful. If any concerns remain, we would be truly grateful for even a brief comment, and we would very much welcome the opportunity to respond further or provide any additional clarification before the discussion closes.

Thank you once again for your valuable time and thoughtful feedback.

Respectfully,
The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Supporting Experiments (1/2)

Table R1. Refiner comparisons

Each refiner was trained for 10,000 steps on DL3DV. Evaluation: 140 DL3DV scenes, 224×224, PSNR (dB). Baseline rows give absolute PSNR; “+” rows give changes from that baseline. Bold numbers mark the highest refiner gain per backbone, camera regime, and view count, including ties.

AnySplat values were not reported for (a) or (b); its available results are included in (c).

Table R1(a). GT poses + GT intrinsics

Backbone / added refiner2v4v6v12v24v
YoNoSplat18.7923.3024.5223.5022.94
+ ReSplat+0.05+0.03+0.05-0.22-0.19
+ IFSplat+0.07+0.26+0.36+0.07+0.05
+ GIFSplat+0.23+0.37+0.46+0.12+0.10
+ DeltaSplat+0.19+0.70+1.01+0.39+0.30
ReSplat backbone only17.5221.9123.2823.8624.30
+ ReSplat+0.49+0.91+0.95+0.61+0.50
+ IFSplat+0.28+0.42+0.56+0.40+0.36
+ GIFSplat+0.44+0.63+0.66+0.45+0.40
+ DeltaSplat+0.37+0.48+0.61+0.54+0.42
ReSplat + ReSplat’s refiner (pretrained)18.8424.6326.3126.7227.12
+ IFSplat+0.06+0.10+0.15+0.13+0.16
+ GIFSplat+0.15+0.16+0.20+0.15+0.18
+ DeltaSplat+0.12+0.30+0.40+0.50+0.58

Table R1(b). Predicted poses + GT intrinsics

Backbone / added refiner2v4v6v12v24v
YoNoSplat19.4323.5424.7423.5122.64
+ ReSplat+0.00-0.01-0.01-0.10+0.02
+ IFSplat+0.01+0.24+0.30+0.11+0.16
+ GIFSplat+0.15+0.37+0.39+0.16+0.21
+ DeltaSplat+0.15+0.59+0.88+0.48+0.63
ReSplat backbone only18.0921.9623.2923.6523.77
+ ReSplat+0.48+0.90+0.93+0.57+0.44
+ IFSplat+0.29+0.43+0.54+0.32+0.24
+ GIFSplat+0.43+0.56+0.63+0.44+0.34
+ DeltaSplat+0.35+0.44+0.57+0.34+0.25
ReSplat + ReSplat’s refiner (pretrained)19.4324.5826.1926.2326.15
+ IFSplat+0.05+0.13+0.14+0.14+0.16
+ GIFSplat+0.11+0.20+0.18+0.17+0.19
+ DeltaSplat+0.17+0.29+0.41+0.51+0.56

Table R1(c). Predicted poses + predicted intrinsics

Backbone / added refiner2v4v6v12v24v
YoNoSplat19.3423.2424.3523.2522.38
+ ReSplat+0.00-0.04-0.03-0.14+0.00
+ IFSplat+0.00+0.11+0.18+0.04+0.08
+ GIFSplat+0.15+0.22+0.27+0.11+0.12
+ DeltaSplat+0.07+0.39+0.58+0.24+0.48
AnySplat15.1119.8221.6620.9520.62
+ ReSplat+1.27+0.89+0.54+0.07-0.25
+ IFSplat+0.52+0.86+0.37+0.40+0.35
+ GIFSplat+0.67+0.92+0.46+0.46+0.38
+ DeltaSplat+0.47+0.79+1.05+1.25+1.35
ReSplat backbone only17.9821.6522.8623.2123.27
+ ReSplat+0.47+0.82+0.82+0.51+0.38
+ IFSplat+0.29+0.36+0.47+0.36+0.27
+ GIFSplat+0.45+0.55+0.56+0.42+0.30
+ DeltaSplat+0.30+0.41+0.52+0.29+0.22
ReSplat + ReSplat’s refiner (pretrained)19.2223.9525.2925.3825.20
+ IFSplat+0.07+0.07+0.11+0.13+0.11
+ GIFSplat+0.15+0.12+0.16+0.17+0.12
+ DeltaSplat+0.15+0.25+0.34+0.42+0.47
AuthorsOpenReview ↗

Supporting Experiments (2/2)

Table R2. Comparison in the ReSplat main-table setting

DL3DV, 140 scenes, 256 × 448. Entries are absolute PSNR (dB). The columns denote input-view counts.

Reconstruction / refinement8 views16 views32 views
ReSplat backbone (K = 0)27.3727.0526.30
+ DeltaSplat refiner (K = 3)27.7327.4426.72
+ ReSplat refiner (K = 4)29.2029.0128.30
ReSplat (K = 4) + DeltaSplat (K = 3)29.5029.3728.68

Table R3. Reported refiner execution time (seconds)

Execution times for refinement at different numbers of input views.

Refiner2v4v6v8v12v24v
ReSplat0.170.200.200.220.280.49
IFSplat0.100.100.150.200.300.70
GIFSplat0.400.701.101.402.104.30
DeltaSplat0.080.130.160.230.360.84

Table R4. Results without evaluation-time pose alignment

DL3DV, 224 × 224; target views are rendered at dataset-provided extrinsics. All methods use predicted context poses.

Intrinsics / viewsYoNoSplat PSNRDeltaSplat PSNRYoNoSplat LPIPSDeltaSplat LPIPS
Known / 623.3224.040.1560.139
Known / 1220.9521.360.2100.197
Known / 2420.2220.680.2350.219
Predicted / 622.2922.410.1730.165
Predicted / 1220.3820.450.2290.217
Predicted / 2419.7119.810.2550.240

Table R5. Test-time refinement iterations

DL3DV, six context views, predicted poses and known intrinsics. The model is trained with three shared-weight iterations. Times include backbone reconstruction.

Iterations K0123456
PSNR ↑24.8026.0226.4526.6426.4826.3226.17
SSIM ↑0.8180.8470.8570.8620.8580.8530.848
LPIPS ↓0.1400.1260.1180.1150.1190.1240.130
Time (ms) ↓5490125160194230266

Table R6. Main results across camera regimes

DL3DV, 140 scenes, 224 × 224; absolute PSNR (dB). These are the original manuscript results, separate from the additional 10,000-step comparison in Table R1.

Method / supplied camera information6v12v24v
YoNoSplat · GT pose + GT intrinsics24.7223.2922.67
DeltaSplat · GT pose + GT intrinsics26.7025.3324.72
YoNoSplat · predicted pose + GT intrinsics24.8923.1522.35
DeltaSplat · predicted pose + GT intrinsics26.6425.0524.21
YoNoSplat · predicted pose + predicted intrinsics24.5322.9322.17
DeltaSplat · predicted pose + predicted intrinsics25.6924.3423.66

Table R7. Expanded component ablation

Higher PSNR and SSIM are better; lower LPIPS is better. The depth-only and ray-only variants isolate the contributions of the two geometric inputs.

VariantPSNR ↑SSIM ↑LPIPS ↓
Full model: rays + rendered depth26.640.8620.115
Without geometric conditioning25.270.8230.146
Depth only (rays removed)26.250.8420.124
Ray only (depth removed)25.930.8360.135
Single-branch mixer26.050.8460.132
Single MLP instead of per-attribute heads26.220.8510.127

Table R8. Rendering quality under camera perturbations

DL3DV, 140 scenes, 224 × 224. Each entry is ReSplat / DeltaSplat PSNR (dB). Columns show rotation / translation perturbation levels.

Input views1° / 2%2° / 5%5° / 10%
218.11 / 19.4617.19 / 18.8516.23 / 18.14
419.32 / 19.9317.50 / 18.1115.80 / 16.43
619.75 / 20.5317.79 / 18.4316.13 / 16.65
1218.29 / 19.0816.44 / 16.9815.24 / 15.45
2418.35 / 18.9916.39 / 16.9115.32 / 15.45

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.